Neural rendering enables dynamic tomography

X-ray computed tomography (X-CT) is a prominent method to observe the internal composition of objects.It requires solving the inverse problem of reconstructing the 3d object from 2d X-ray projections.Thousands of projections are typically needed to obtain a detailed reconstruction.Therefore, it has never been possible to reconstruct material deformation under dynamic high-speed loading, which has impeded our understanding of high-speed deformation processes.In this work, we address this limitation by combining high-fidelity X-CT with differentiable neural rendering.In our two-stage approach, we first reconstruct the canonical volume and then use a neural network to predict a temporal deformation field with a cubic spline output parametrization. We demonstrate the reconstruction of deforming objects from very few (two) projections which enables a paradigm shift in reconstruction of dynamic experiments.